Boom5426 commited on
Commit
433f214
·
verified ·
1 Parent(s): ca00568

Upload 更新.md

Browse files
Files changed (1) hide show
  1. code/更新.md +137 -0
code/更新.md ADDED
@@ -0,0 +1,137 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 新题目:
2
+
3
+ VCDesign: Finite-Budget Intervention Design for Virtual Cells
4
+
5
+
6
+
7
+ 摘要:
8
+
9
+ Virtual-cell models are typically judged by how well they predict the effects of perturbations or recover the interventions associated with a cellular transition. Yet experimental planning poses a different question: given a desired cellular state, which interventions should be tested when only a limited number of experiments can be performed? Here we introduce VCDesign, which, to our knowledge, is the first framework to formulate target-conditioned cellular intervention design as finite-budget ranking over a variable candidate set. Given source and target states, a candidate library and an experimental budget, VCDesign prioritizes the interventions to test. We also introduce a decision-aligned evaluation framework that measures the quality of the selected candidates using independently measured held-out outcomes. For candidates lacking historical perturbation responses, we develop VCDesign-CED, which uses existing perturbation measurements and biological knowledge to infer candidate effects and combine them with direct candidate scoring. Across four CRISPRi Perturb-seq contexts, VCDesign-CED improves the prioritization of response-unseen candidates, increasing pooled budgeted utility at 20 selections by 0.049 and recovering 1.11 additional top-5% candidates on average. When candidate responses are already available in other contexts, direct retrieval performs better, and routing candidates according to response availability further improves mixed-library selection. Controlled analyses further show that transferable perturbation evidence and candidate knowledge contribute more to design performance than simply adding training contexts or increasing effect-model complexity, while prediction accuracy does not reliably rank nearby models by their utility for intervention design. Together, these results establish a task and evaluation framework for using virtual-cell models to make experimental decisions under finite budgets.
10
+
11
+
12
+ 引言:
13
+
14
+ Virtual-cell models aim to predict how cells respond to perturbations and, ultimately, to help identify interventions that can move cells towards desired states. Most work has focused on two neighbouring problems: forward models predict the cellular consequences of a specified genetic, chemical or other perturbation, whereas inverse models start from a desired transition and seek interventions that might produce it. Experimental planning requires a further step, because researchers typically face a starting state, a desired target state, a set of interventions that can actually be tested and a budget that allows only a small fraction of them to be evaluated. The practical question is therefore not only what a perturbation will do, or which intervention best matches an observed transition, but which interventions should be tested first when experimental resources are limited. Framing cellular intervention design in these terms turns it into a decision problem in its own right.
15
+
16
+ Several lines of work already move towards this goal, but they approach it through different computational routes. Expression-profile resources such as Connectivity Map and LINCS retrieve perturbations whose measured responses match a desired transcriptional change; inverse models such as CellNavi predict candidate drivers directly from source and target states; forward models can simulate candidate interventions and rank their predicted outcomes; and more recent approaches search for interventions through generative or optimization procedures. Although these methods differ in how they represent interventions, what information they require and what objective they optimize, they ultimately serve the same experimental decision: which candidates in the available library should occupy the limited number of positions that can actually be tested? Existing evaluation does not always align with this decision, because prediction error measures the quality of a response estimate and identity recovery asks whether a method retrieves the intervention associated with a query transition, whereas an experiment executes only a small prefix of a ranking. A useful design system should also reward multiple distinct interventions when they produce similarly valuable outcomes, making a common task and evaluation interface necessary for comparing different approaches on the basis of the decisions they enable.
17
+
18
+ Here we introduce VCDesign, which formalizes this decision by asking a model to rank a feasible intervention library given source and target states and an experimental budget \(B\), and to prioritize the \(B\) candidates to test. To our knowledge, this is the first explicit formulation of target-conditioned cellular intervention design as finite-budget ranking over a variable candidate set. We pair this task with a decision-aligned evaluation framework that focuses on the portion of the ranking that can actually be executed and assigns value to selected candidates using independently measured held-out outcomes, thereby allowing multiple interventions to be rewarded when they produce useful effects and placing retrieval, forward simulation, optimization and direct scoring under a common downstream objective. VCDesign also makes information availability part of the problem definition: generalizing to a new cellular context is different from ranking a candidate whose own perturbation response has never been available for effect learning. For this response-unseen setting, we develop VCDesign-CED, which uses historical perturbation measurements to learn how deployment-available biological knowledge relates to perturbational effects and then uses the inferred effects to help rank candidates for a requested cellular transition. In this way, experimental evidence from previously measured interventions can inform the prioritization of candidates that have not themselves been profiled.
19
+
20
+ We evaluate this framework across four CRISPRi Perturb-seq contexts and find that historical perturbation evidence consistently improves finite-budget selection for response-unseen candidates. Across three held-out cellular contexts, VCDesign-CED increases pooled BU@20 by 0.049 over the same base ranker and identifies, on average, 1.11 additional candidates from the top 5% of measured utility within the first 20 selections, with the same pattern also observed for held-out intervention identities in K562. The experiments further reveal a clear boundary set by information availability: when a candidate has already been measured in other cellular contexts, direct retrieval of those responses outperforms effect inference, whereas routing measured and unmeasured candidates to different solvers further improves selection in mixed libraries. Controlled analyses also show that response-unseen design depends more strongly on transferable perturbation evidence and informative candidate knowledge than on simply adding training contexts or increasing effect-model complexity; increasing the amount of measured evidence improves design utility, whereas additional contexts at matched measurement volume and the richer effect models we tested provide little or no comparable benefit. Prediction accuracy captures some large differences across data and model configurations, but it does not reliably order nearby models by their utility for intervention design. Together, these results extend virtual-cell modelling from describing cellular states towards supporting experimental decisions, and provide a common task and evaluation foundation for prioritizing high-value interventions under finite experimental budgets.
21
+
22
+ Our contributions are threefold:
23
+
24
+ A decision formulation for cellular intervention design. We define intervention design as selecting the most promising candidates from a variable, feasible intervention set under a fixed experimental budget, separating the experimental objective from the computational method used to produce the ranking.
25
+
26
+ An evaluation framework aligned with experimental choice. We evaluate the candidates that would actually be tested and judge them by independently measured held-out outcomes, allowing different computational approaches to be compared by the quality of the experimental decisions they support.
27
+
28
+ A strategy for interventions without measured responses. We introduce VCDesign-CED to transfer information from previously measured perturbations through biological knowledge to prioritize response-unseen candidates, and use controlled analyses to determine when this transfer succeeds and which sources of information matter most.
29
+
30
+
31
+ ### Discussion and conclusion
32
+
33
+ By making **experimental choice itself** the common object of evaluation, VCDesign reveals a principle that is more fundamental than the performance of any single model: **the ability to design cellular interventions is not an intrinsic property of a model alone, but depends on what experimental evidence is available and how effectively that evidence can be converted into decisions under a finite budget.** Once different computational approaches are evaluated on this common decision scale, a clear information hierarchy emerges. When historical responses for a candidate are already available, those measurements provide the strongest basis for prioritization; when they are not, perturbation atlases retain value beyond the interventions they directly contain, because evidence accumulated from other perturbations can be transferred through biological knowledge to improve the ranking of response-unseen candidates across both held-out identities and cellular contexts. Mixed candidate libraries reinforce this view: intervention design is not a search for a single model that should dominate every regime, but a problem of matching the available evidence to the appropriate computational route. Controlled analyses further suggest that the remaining bottleneck in response-unseen design lies primarily in the **transferable evidence itself**. Expanding perturbation measurements and providing informative candidate knowledge materially improve design quality, whereas adding training contexts alone or increasing effect-model complexity does not yield comparable gains; prediction accuracy likewise does not reliably determine which nearby models will make better experimental choices. Moving from prediction and identification to design therefore changes more than the required model output—it changes the scientific question to **how biological knowledge and experimental measurements can be organized into useful evidence for deciding what to test next**.
34
+
35
+ These conclusions remain bounded by the scope of the present evidence. Our experiments are restricted to single-gene CRISPRi perturbations across four cell lines and primarily establish improvements in transcriptomic directional prioritization; the realized-outcome analyses do not yet show that these ranking gains translate consistently into improved biological endpoints. Response-unseen transfer also weakens as candidate libraries approach genome scale, indicating that the value of transferable evidence does not automatically extend to arbitrarily large search spaces. At the same time, the data contain substantial and reproducible context-specific perturbation structure, yet the limited number of training contexts is insufficient to learn a context-conditioned effect model that can exploit this structure reliably. This points to a broader challenge for the field: progress may depend first on **what should be measured, where it should be measured and how experimental coverage should be expanded**, and only then on how increasingly powerful models organize that evidence. Addressing this challenge will require perturbation atlases with greater biological breadth and information content, better integration of candidate-specific measurements with model-based inference, and prospective experiments that test whether computational prioritization produces genuine biological benefit. **If virtual cells are to become part of experimental science, their next defining capability will be to turn the growing body of biological measurements into reliable decisions about which experiments are most worth doing next.**
36
+
37
+
38
+
39
+ 其他部分:
40
+
41
+ 收到。这个 `更新.md` 我会作为**当前最新且优先级最高的写作基线**:新题目、摘要、四段引言、三项贡献以及两段 Discussion/Conclusion 都以这里为准,后续正文修改必须服务于这条主线,而不是再反过来改变它。
42
+
43
+ 基于这份最新内容,我会把全文主线冻结成一句话:
44
+
45
+ > **VCDesign establishes cellular intervention design as a finite-budget experimental decision, aligns evaluation with the choices that can actually be made, and reveals that the value of different computational routes depends fundamentally on what perturbation evidence is available.**
46
+
47
+ 这也意味着上一轮说的“确定没问题的正文修改”需要稍微调整。现在我认为可以直接执行、且不会改变任何数据事实的有下面四处。
48
+
49
+ ### 1. Results opening:从 CED-centered 改成整篇 paper-centered
50
+
51
+ 当前:
52
+
53
+ > We ask how well VCDesign-CED selects, when it is the right solver, and what limits it.
54
+
55
+ 已经和新的摘要/引言不匹配。建议改为:
56
+
57
+ > **We use VCDesign to ask how different computational routes support the same finite-budget decision, whether historical perturbation evidence can guide candidates whose own responses are unavailable, and what determines the value of this transfer. We first compare the evaluated solvers under a common decision objective and their respective information requirements, then examine response-unseen transfer, response-dependent solver choice and the factors that limit intervention design.**
58
+
59
+ 这样 **Table 1 自然成为 Results 的第一张核心证据**,而不是 VCDesign-CED 的附属 baseline table。
60
+
61
+ ---
62
+
63
+ ### 2. Table 1:明确它的真正责任
64
+
65
+ Table 1 现在最重要的不是“谁最好”,而是同时展示:
66
+
67
+ **common decision objective + different computational routes + different information requirements**。
68
+
69
+ caption 开头建议改成:
70
+
71
+ > **Table 1: Evaluated solvers under a common finite-budget decision objective and different information requirements.**
72
+
73
+ 正文在进入 Figure 2 之前,加一个高层解释:
74
+
75
+ > **Table 1 reveals a first organizing principle of intervention design: the usefulness of a computational route depends on the information available for each candidate.** Methods in the upper block can rank candidates without requiring candidate-specific measured responses, whereas those in the lower block rely on such measurements. Under the response-unseen setting, VCDesign-CED improves selection over the direct-scoring and forward-then-match alternatives evaluated here; when historical candidate responses are available, direct profile retrieval is substantially stronger. These regimes motivate the analyses below, where we first test whether evidence from other perturbations can support response-unseen candidates and then examine how measured and unmeasured candidates should be handled when they coexist.
76
+
77
+ 这段和你现在 Discussion 的 **information hierarchy** 完全呼应,但仍严格受 Table 1 数据支持,没有把它写成 universal leaderboard。
78
+
79
+ ---
80
+
81
+ ### 3. Section 3.2:把 Evaluation contribution 真正立起来
82
+
83
+ 现在新的 contribution bullet 已经明确把 evaluation 独立成第二贡献,因此正文 3.2 也必须有同样高度。
84
+
85
+ 建议在具体 CRISPRi cosine 之前加:
86
+
87
+ > **Evaluation should reflect the experimental choice that the model is intended to support.** Under a finite budget, only the highest-ranked candidates can be tested, so the relevant object is the quality of this executable ranking prefix rather than the full ordering. Candidate value should also be determined by independently measured outcomes rather than by recovery of the intervention that generated the query, allowing multiple interventions to be useful for the same desired transition. This separates the downstream decision from the mechanism used to produce the ranking and allows retrieval, forward simulation, optimization and direct scoring to be evaluated through a common decision objective. **We instantiate these principles for CRISPRi as follows.**
88
+
89
+ 然后再接现有:
90
+
91
+ > In our CRISPRi benchmark, useful interventions reproduce...
92
+
93
+ 这样 Equation 1 和 BU@B 就变成 **evaluation principles 的一个实例**,而不是贡献本身。
94
+
95
+ ---
96
+
97
+ ### 4. Section 4 opening:CED 明确只是 response-unseen 的实例
98
+
99
+ 新的三项贡献已经写得很清楚:第三项是
100
+
101
+ > **A strategy for interventions without measured responses.**
102
+
103
+ 因此 Section 4 不应让读者误以为整篇论文的核心是某个 architecture。
104
+
105
+ 建议开头改成:
106
+
107
+ > **VCDesign-CED instantiates the framework for the response-unseen setting, where a candidate’s own perturbation response is unavailable for effect learning.** The challenge is to use evidence accumulated from profiled interventions without assuming access to the candidate’s own measured response. VCDesign-CED combines direct candidate scoring with Candidate Effect Distillation (CED), which learns how deployment-available biological knowledge relates to measured perturbational effects. The inferred effect is then aligned with the requested cellular transition and used as complementary evidence for ranking.
108
+
109
+ 这里我会坚持使用:
110
+
111
+ > **response unavailable for effect learning**
112
+
113
+ 而不是泛化成:
114
+
115
+ > unseen / unprofiled candidate
116
+
117
+ 因为后者对所有实验设置并不严格成立。
118
+
119
+ ---
120
+
121
+ ## 目前我反而不建议继续改的地方
122
+
123
+ **摘要、引言和 Discussion/Conclusion 暂时停止润色。** 这三部分现在已经高度一致,继续来回改很容易边际收益转负。尤其 Discussion 第一段现在已经把 **Table 1 → Figures 2–4** 汇总成 “information hierarchy / evidence-to-decision” 这条科学发现,这是正确的最终落点。
124
+
125
+ **Figure 2–4 的 caption 也先不动。** 当前 caption 对数据的描述总体准确,尤其 Figure 3 和 Figure 4 已经是结论式 caption。
126
+
127
+ **5.4 暂时不要进一步强化成 `evidence is the bottleneck`。** 你的 Discussion 可以做这种综合解释,但 Results 仍应保留“measurement volume、identity coverage、context count 尚未完全解耦”的事实边界。否则会出现首尾严谨、正文过度归因的问题。
128
+
129
+ ---
130
+
131
+ 还有一个小地方我建议以后统一:题目现在的
132
+
133
+ > **VCDesign: Finite-Budget Intervention Design for Virtual Cells**
134
+
135
+ 我支持保留。它比 `Candidate-Conditioned Inverse Modeling...` 更准确,因为现在论文的主体已经明确是 **design problem + evaluation + information regimes**,而不是 inverse architecture。
136
+
137
+ 所以接下来最合理的动作不是继续动首尾,而是**按这份最新 narrative contract,逐节改 Section 3–5,让正文真正和新的摘要、引言、结论成为同一篇 paper。**